Clean Label Poisoning Attacks: From Classification to Speech Recognition

Explore the world of poisoning attacks and defenses in speech recognition through this 16-minute video presentation by Henry Li Xinyuan from the Center for Language & Speech Processing at JHU. Delve into adversarial attacks and their ability to manipulate neural networks, focusing on the emerging threat of poisoning attacks that compromise model integrity through training data manipulation. Examine different attack strategies, including dirty and clean label attacks, and learn about innovative defense mechanisms like DINO-based cluster-and-filter defenses. Gain insights into cybersecurity, machine learning, and the ongoing battle between AI advancements and adversarial threats. Unpack complex topics, discuss potential defenses, evaluate the efficacy of various strategies, and explore future research directions in this engaging presentation.

Clean Label Poisoning Attacks: from Classification to Speech Recognition